Flux 3 first-last vs Veo first-last

Flux 3 First Last Frame to Video is 34 credits, image-to-video, native audio, 5–20 seconds, two stills as endpoints. VEO First Last Frame is 64 credits, image-to-video, native audio, 4K, 4 / 6 / 8 seconds, two stills as endpoints. Pick Flux when the job is a longer interpolation. Pick Veo when the job is a short 4K take with Veo’s dialogue and physics. Duration versus 4K dialogue. Both rows need frames. Neither row is text-to-video.

This is a job pick. First-last is a contract at both ends: frame one is the start still, frame last is the end still, the model fills the travel. The prompt names what happens in between. Re-describing the product in prose is how the endpoints and the sentence fight.

Black Forest Labs positioned FLUX 3 Video as clips up to 20 seconds with native audio (BFL, 4 August 2026). Early evals BFL published on 23 July 2026 preferred Flux 3 over several peers; those are BFL’s results, not Versely benchmarks. The catalog sticker here is 34 credits.

Veo’s first-last card is Google’s I2V-adjacent Veo row: 4K tier, native audio, 4 / 6 / 8 seconds, 64 credits. The model does not change because you launched it from Flow. Google Flow is a Veo surface, not a studio. This page is the row, not the launcher.

Flux 3: 34 credits, 5–20 seconds

The Flux card: audio true, category image-to-video, image required, durations every second from 5s through 20s, credits 34. The description: interpolates a smooth, coherent transition between a defined start frame and end frame, with native audio, up to 1080p, 5–20 seconds. No 4K token on the card. Do not write 4K into a Flux brief as if the field were filled.

Thirty-four credits is the catalog sticker for a 5-second clip. Flux bills per second, at 7 credits a second on the card, and a 20-second clip tops out at 136.

When Flux is the pick:

You have a start still and an end still of the same SKU, same talent, or same room, and you need more than eight seconds of travel.
The feed is 1080p or below. A 4K insertion order is not this row.
You want native audio on a long hold — a pour, a walk, a lid, a curtain — without paying 64.
Look-dev: you will throw the take away if the interpolation is wrong. Thirty-four credits is how you find out whether the two stills can coexist.

When Flux is the wrong pick:

You need 4K.
You need Veo’s always-on dialogue in a cinematic eight.
You have only one still. First-last needs two. One still is I2V on a different row: Wan 2.6 Image to Video at 20 credits, or Kling O3 Standard Image to Video at 45.
You have only a sentence. That is T2V. Flux 3 Text to Video is the sentence sibling at 34 credits, 5–20s, native audio — not this comparison.

Native audio is on. Write whether anyone speaks. A 20-second surprise monologue is a long, cheap error. SKU interpolations usually want foley and no voice. Say “no speech.”

The two stills have to be a pair. Same crop family, same light, same pack. A studio packshot as frame one and a phone grab as frame last is how you get a bottle that is in neither photo. The end still is not a new product. If the end still is a different SKU, you advertised a morph.

Veo: 64 credits, 4K, 4 / 6 / 8 seconds

The Veo first-last card: audio true, image-to-video, image required, 4 / 6 / 8 seconds, 4K, 64 credits. Description: generate videos between first and last frames. Same job shape as Flux. Different ceiling. Different sticker. Different dialogue.

Sixty-four credits is the catalog sticker, and a full job runs 64 to 256 depending on audio and length. It is the same sticker as Veo 3.1 T2V. You are not paying a first-last surcharge on this card. You are paying Veo. The meter is 4K and the 4 / 6 / 8 menu, not a 20-second hold.

When Veo first-last is the pick:

The insertion order is 4K, or you will crop a 4K plate into 9:16 and still want the detail.
Someone speaks, and you want Veo’s native audio on that line, in eight seconds or less.
The travel is a short cinematic: a camera move between two locked frames, a product that must land on a specific end pose, a talent who must arrive at a specific expression.
You already know the stills. Look-dev on 64 credits is how a first-last model becomes a sketchbook. Draft the pair on Flux if 1080p is enough to learn. Promote to Veo when 4K dialogue is the ship.

When Veo first-last is the wrong pick:

You need 10, 15, or 20 seconds of travel. Veo will not grow. Prompting “continue” does not add a slider.
You wanted reference-to-video: identity in a new scene, not interpolation between two frames. That is VEO 3.1 Reference to Video at 128 credits, 8 seconds only, up to three stills. First-last is not a stack. The stack is not a first frame.
You wanted mute 20-credit motion. That is Wan 2.6 Image to Video Flash.

Google made the visible watermark optional on Gemini app and Flow on 14 August 2026 (Fastlane, 24 August 2026). SynthID and C2PA stay embedded. Optional chrome is not a reason to treat Flow as the studio. Keep Veo. Leave the surface if the slate is not only Veo.

Duration versus 4K dialogue

The honest table is short.

Need 9–20 seconds. Flux. Veo cannot.
Need 4K. Veo. Flux’s card says up to 1080p.
Need both 20 seconds and 4K. Neither row. Split the job, or change families. Do not prompt Flux as if 4K were a quality token. Do not prompt Veo as if 20s were a sentence.
Need 4, 6, or 8 seconds at 1080p with native audio. Either can run. Veo is 64 credits and 4K you may not use. Flux is 34 credits starting at 5s — there is no 4s on Flux. A 4-second Veo take is a Veo-only length.
Need a talking presenter in a short cinematic. Veo, if the stills are endpoints. If you have no end still, Veo T2V or a talking-head row, not first-last.

Sora 2 Pro storyboard was 240 credits, I2V, native audio, 10 / 15 / 25 seconds. The Sora API shuts 24 September 2026 (OpenAI Help Centre, September 2026). A leftover storyboard brief that was two keyed frames is closer to Flux’s 5–20s than to Veo’s 8s. It is not a reason to invent a Sora row that still runs.

Article 50 still applies to photoreal people in either interpolate. First-last is not an exemption. Label at first exposure for EU audiences since 2 August 2026.

Captions after the take. Native audio is not typesetting. The AI caption generator is the burn-in. Keep the block off the pack if the pack is the proof.

Production order

Lock both stills before you pick the row. Approve type at 100%. If the end still is a different label, stop.

If you do not know whether the pair interpolates, run Flux at 5s for 34 credits. If the travel holds and you only needed 1080p, stay. If the travel holds and the insertion order is 4K dialogue in 8 seconds or less, rerun the same pair on Veo at 64.

If you already know it is an eight-second 4K line, skip Flux. You will not learn 4K on a 1080p draft. You will only learn whether the stills agree.

The AI video generator is the launcher. Two uploads. One prompt that names the travel, not a second product.

Thirty-four credits is the length door. Sixty-four is the 4K dialogue door. Same job shape. Different ceilings.

Run the pair on the ceiling you actually need in Versely.

FAQ

Can Veo first-last go past 8 seconds?

No. 4, 6, or 8. For 9–20 seconds of interpolation, Flux 3 first-last at 34 credits.

Does Flux 3 first-last output 4K?

The published card says up to 1080p. Do not brief 4K on that slug. Veo first-last is the 4K row at 64 credits.

What if I have only one still?

Neither first-last row will do the job correctly. Use a single-still I2V row, or make the end still before you interpolate.